feat: Cohere chat models integration

[tcapelle]

Cohere API integration with Weave

[circle-job-mirror[bot]]

Preview this PR with FeatureBee: https://beta.wandb.ai/?betaVersion=f1a7d7bc2d4f8f59423385562763157622dd9059

[tcapelle]

@tssweeney I need help with 2 things:

• the output of the cohere SDK is a pydantic.v1.BaseModel that we don't parse correctly
• I think the wrapped function in cohere SDK patches after the weave patch, and that's why the name is ignored. For the streaming version the naming is correctly displayed, I suppose the accumulator fixes the naming.

[tcapelle]

The only thing that may need changing is how the patching is applied.

As you know @tssweeney the cohere Client modifies the chat function on the init, so we need to patch before creating a client. Two options for this:

• We call weave.init before creating the client (this is weird in general and wouldn't be how most of of the users do, but... it is all over our docs: anthropic, mistral, etc...)
• We manually patch before creating a client. Something closer to where the imports happen. In this PR we do:

import cohere
import os
import weave
from weave.integrations.cohere import cohere_patcher

# we need to patch before we create the client
cohere_patcher.attempt_patch()

co = cohere.Client(api_key=os.environ["COHERE_API_KEY"])

weave.init("cohere_project")

one option could be:

# cohere.__init__.py
from weave.integrations.cohere.cohere_sdk import cohere_patcher as _cohere_patcher

def cohere_patcher():
    _cohere_patcher.attempt_patch()

but feel free to change this import logic...